What are the challenges of accurately detecting sarcasm and context in sentiment analysis models?